earthquake forecasting: statistics and information
TRANSCRIPT
arX
iv:1
310.
7173
v1 [
phys
ics.
geo-
ph]
27
Oct
201
3
Noname manuscript No.
(will be inserted by the editor)
Earthquake forecasting: Statistics and Information
V.Gertsik , M.Kelbert, A.Krichevets
the date of receipt and acceptance should be inserted later
Abstract We present an axiomatic approach to earthquake forecasting in terms of
multi-component random fields on a lattice. This approach provides a method for con-
structing point estimates and confidence intervals for conditional probabilities of strong
earthquakes under conditions on the levels of precursors. Also, it provides an approach
for setting multilevel alarm system and hypothesis testing for binary alarms. We use
a method of comparison for different earthquake forecasts in terms of the increase of
Shannon information. ’Forecasting’ and ’prediction’ of earthquakes are equivalent in
this approach.
1 Introduction
The methodology of selecting and processing of relevant information about the future
occurrence of potentially damaging earthquakes has reached a reasonable level of matu-
rity over the recent years. However, the problem as a whole still lacks a comprehensive
and generally accepted solution. Further efforts for optimization of the methodology of
forecasting would be productive and well-justified.
A comprehensive review of the modern earthquake forecasting state of knowledge
and guidelines for utilization can be found in [Jordan et all., 2011]. Note that all
methods of evaluating the probabilities of earthquakes are based on a combination
of geophysical, geological and probabilistic models and considerations. Even the best
and very detailed models used in practice are in fact only ’caricatures’ of immensely
complicated real processes.
A mathematical toolkit for earthquake forecasting is well presented in the paper
[Harte and Vere-Jones, 2005]. This work is based on the modeling of earthquake se-
Institute of Earthquake Prediction Theory and Mathematical Geophysics RAS, Moscow, RF,[email protected]
Dept. of Mathematics, Swansea University, Singleton Park, Swansea, SA2 8PP, UK. Institute ofEarthquake Prediction Theory and Mathematical Geophysics RAS, [email protected]
Lomonosov MSU, Department of Psychology, Moscow, Russia, [email protected]
Address(es) of author(s) should be given
2 V.Gertsik, M.Kelbert, A.Krichevets
quences in terms of the marked point processes. However, the mathematical technique
used is quite sophisticated and does not provide direct practical tools to investigate
the relations of the structure of temporal-spatial random fields of precursors to the
appearance of strong earthquakes.
The use of the multicomponent lattice models (instead of marked point processes)
gives a different/novel way of investigating these relations in a more elementary way.
Discretization of space and time allows us to separate the problem in question into
two separate tasks. The first task is the selection of relevant precursors, i.e., observable
and theoretically explained physical and geological facts which are casually related to a
high probability of strong earthquakes. Particularly, this task involves the development
of models of seismic events and computing probabilities of strong earthquakes in the
framework of these models. Such probabilities are used as precursors in the second
task.
The second task is the development of methodology of working with these precur-
sors in order to extract the maximum information about the probabilities of strong
earthquakes. This is the main topic of this paper.
Our approach allows us to obtain the following results:
• Estimates of probabilities of strong earthquakes for given values of precursors are
calculated in terms of the frequencies of historic data.
• Confidence intervals are also constructed to provide reasonable bounds of preci-
sion for point estimates.
• Methods of predictions (i.e., binary alarm announcement [Keilis-Borok, 1996],
[Keilis-Borok, Kossobokov, 1990], [Holiday et all., 2007]) and forecasting (i.e., calcu-
lating probabilities of earthquakes [Jordan et all., 2011], [Kagan and Jackson, 2000],
[Harte and Vere-Jones, 2005], [WGNCEP]) are equivalent in the following sense: the
setting of some threshold for probability of earthquakes allows to update the alarm
level. On the other hand, the knowledge of the alarm domain based on historical data
allows us to evaluate the probabilities of earthquakes. In a sense, the prediction is
equivalent to hypothesis testing as well, see Section 11.
• In our scheme we propose a scalar statistic which is the ratio of actual increment
of information to the maximal possible increment of information. This statistic allows
us to linearly order all possible forecasting algorithms. Nowadays the final judgement
about the quality of earthquake forecasting algorithms is left to experts. This arrange-
ment puts the problem outside the scope of natural sciences which are trying to avoid
subjective judgements.
The foundation of our proposed scheme is the assumption that the seismic process
is random and cannot be described by a purely deterministic model. Indeed, if the
seismic process is deterministic then the inaccuracy of the forecast could be explained
by the non-completeness of our knowledge about the seismic events and non-precision
of the available information. This may explain, at least in principle, attacks from the
authorities addressed to geophysicists who failed to predict a damaging earthquake.
However, these attacks have no grounds if one accepts that the seismic process is
random. At the end of the last century (February-April 1999) a group of leading seis-
mologists organized a debate via the web to form a collective opinion of the scientific
community on the topic: ’Is the reliable prediction of individual earthquakes a realistic
scientic goal?’ (see http://www.nature.com/nature/debates/earthquake/).
Despite a considerable divergence in peripheral issues all experts taking part in the
debate agreed on the following main principles:
Earthquake forecasting: Statistics and Information 3
• the deterministic prediction of an individual earthquake, within sufficiently nar-
row limits to allow a planned evacuation programme, is an unrealistic goal;
• forecasting of at least some forms of time-dependent seismic hazard can be jus-
tified on both physical and observational grounds.
The following facts form the basis of our agreement with this point of view.
The string-block Burridge-Knopov model, generally accepted as a mathematical
tool to demonstrate the power-like Gutenberg-Richter relationship between the mag-
nitude and the number of earthquakes, involves the generators of chaotic behaviour or
dynamic stochasticity. In fact, the nonlinearity makes the seismic processes stochastic:
a small change in the shift force may lead to completely different consequences. If the
force is below the threshold of static friction the block is immovable, if the force exceeds
this threshold it starts moving, producing an avalanche of unpredictable size.
This mechanism is widespread in the Earth. Suppose that the front propagation of
the earthquake approaches a region of enhanced strength of the rocks. The earthquake
magnitude depends on whether this region will be destroyed or remains intact. In the
first case the front moves further on, in the second case the earthquake remains local-
ized. So, if the strength of the rocks is below the threshold the first scenario prevails, if
it is above the threshold the second scenario is adapted. The whole situation is usually
labelled as a butterfly effect : infinitesimally small changes of strength and stress lead to
macroscopic consequences which cannot be predicted because this infinitesimal change
is below any precision of the measurement. For these reasons determinism of seismic
processes looks more doubtful than stochasticity.
The only comment we would like to contribute to this discussion is that the fore-
casting algorithms based exclusively on the empirical data without consistent physical
models could hardly be effective in practice (see Sections 12, 13 for more details).
In conclusion we discuss the problem of precursor selection and present a theorem
by A. Krichevetz stating that using a learning sample for an arbitrary feature selection
in pattern recognition is useless in principle.
Finally, note that our approach may be well-applicable for the space-time forecast-
ing of different extremal events outside the scope of earthquake prediction.
2 Events and precursors on the lattice
In order to define explicitly estimates of probabilities of strong earthquakes we dis-
cretize the two-dimensional physical space and time, i.e., introduce a partition of three-
dimensional space-time into rectangular cells with the space partition in the shape of
squares and time partition in the shape of intervals. These cells should not intersect to
avoid an ambiguity in computing the frequencies for each cell. In fact, the space cells
should not be perfect squares because of the curvature of the earth’s surface, but this
may be neglected if the region of forecasting is not too large.
So, we obtain a discrete set ΩK with N = I × J × K points which is defined as
follows. Let us select a rectangular domain A of the two-dimensional lattice with I ×J
points x = (xi, yj); xi = a×i; i = 1, . . . , I and yj = a×j; j = 1, . . . , J,a is the step of
the lattice. A cell in ΩK takes the shape of parallelepiped of height Δt with a square
base. Clearly, any point in ΩK has coordinates (xi, yj , tk), tk = t+k∆t; k = 0, . . . ,K.
.
We say that a seismic event happens if an earthquake with magnitude greater than
some pre-selected threshold M0 is registered, and this earthquake is not foreshock or
4 V.Gertsik, M.Kelbert, A.Krichevets
aftershock of another, more powerful earthquake (we put aside a technical problem of
identification of foreshocks and aftershocks in the sequence of a seismic event). For any
cell in our space-time grid we define an indicator of an event, i.e., a binary function
h(i, j, k). This function takes the value 1 if at least one seismic event is registered
in a given cell and 0 otherwise. Suppose that for all points (xi, yj , tk) the value of a
vector precursor f(i, j, k) = fq(i, j, k), q = 1, ..., Q is given. The components of the
precursor fq(i, j, k), q = 1, ..., Q are the scalar statistics constructed on the base of our
understanding of the phenomena that precede a seismic event.
Remark 1 Note that specifying an alarm domain as a circle with center at a lattice
site and radius proportional to the maximal magnitude of the forecasted earthquakes
leads to a contradiction. Indeed, suppose we announce an alarm for earthquakes with
minimal magnitude 6 in a domain A6. Obviously, the same alarm should be announced
in the domain A7 as well. By the definition A6 ⊂ A7 and we expect an earthquake
with magnitude at least 7 and do not expect an earthquake with magnitude at least 6in the domain A7 \ A6. But this is absurd.
3 Mathematical assumptions
A number of basic assumptions form the foundation of the mathematical tecnique of
earthquake forecasting. In the framework of mathematical theory they can be treated as
axioms but are, in fact, an idealization and simplification with respect to the description
of the real phenomena. Below we summarize the basic assumptions which are routinely
used in existing studies of seismicity and algorithms of earthquake forecasting even the
authors do not always formulate them explicitly.
We accept the following assumptions or axioms of the mathematical theory:
(i) The multicomponent random process h(i, j, k), f(i, j, k), describing the joint evo-
lution of the vector precursors and the indicator of seismic events, is stationary.
This assumption provides an opportunity to investigate the intrinsic relations be-
tween the precursors and the seismic events based on the historical data. In other words,
the experience obtained by analysing the series of events in the past, is applicable to
the future as the properties of the process do not depend on time.
In reality, this assumption holds only approximately and for a restricted time pe-
riod. Indeed, plate tectonics destroys the stationarity for a number of reasons including
some purely geometrical considerations. For instance, the movements of the plates leads
to their collisions, their partial destruction and also changes their shapes. Nevertheless,
the seismic process can be treated as quasi-stationary one for considerable periods of
time. At the time when the system changes one quasi-stationary regime to another (say,
nowadays, many researcher speak about the abrupt climate change) the reliability of
any prediction including the forecast of seismic events is severely restricted.
(ii) The multicomponent random process h(i, j, k), f(i, j, k) is ergodic.
Any quantitative characteristic of seismicity more representative than a registration
of an individual event is, in fact, the result of averaging over time. For instance, the
Gutenberg-Richter law, applied to a given region relates the magnitude with the average
number of earthquakes where the averaging is taken over a specific time interval. In
order to associate with the time averaging a proper probabilistic characteristic of the
process and make a forecast about the future one naturally needs the assumption of
ergodicity. This exactly means that any averaging over time interval [0, T ] will converge
Earthquake forecasting: Statistics and Information 5
to the stochastic average when T → ∞. In view of ergodicity one can also construct
unbiased and consistent estimates of conditional probabilities of strong earthquakes
under conditions that the precursors take values in some intervals. Naturally, these
estimates are the frequencies of observed earthquakes, i.e., ratios of the number of cells
with seismic events and prescribed values of precursors to the total number of cells
with the prescribed values of precursors. (Recall that an unbiased point estimate θ of
parameter θ satisfies the condition Eθ = θ, and a consistent estimate converges to the
true value θ when the sample size tends to infinity).
(iii) Any statement about the value of the indicator of a seismic event h(i, j, k) in the
cell (i, j, k) or its probability should be based on the values of the precursor f(i, j, k)only.
This assumption means that the precursor in the given cell accumulates all the
relevant information about the past and the information about the local properties of
the area that may be used for the forecast of the seismic event in this cell. In other
words, the best possible precursor is used (which is not always the case in practice). As
in the other cases, this assumption is only an approximation to reality, and the quality
of a forecast depends on the quality of the selection and accumulation of relevant
information in the precursors.
Below we present some corollaries and further specifications.
(iii-a) For any k the random variables h(i, j, k), i = 1, ..., I, j = 1, . . . , J are condi-
tionally independent under the condition that the values of any measurable function
u(f(i, j, k)) of the precursors f(i, j, k), i = 1, . . . , I, j = 1, . . . , J are fixed..
In practice this assumption means that the forecast for the time tk = t0 + k∆t
cannot be affected by the values related to the future time intervals (tk, tk + ∆t].In reality all of these events may be dependent, but our forecast does not use the
information from the future after tk.
(iii-b) The conditional distribution of the random variable h at a given cell depends
on the values of the precursors at this cell f and is independent of all other variables.
(iii-c) The conditional probabilities Prij h | u(f) of the indicator of seismic events h
in the cell (i, j, k), under condition u(f) in this cell do not depend on the position of
the cell in space (the time index k related to this probablity may be dropped due to
the stationarity of the process).
In other words, the rule for computing the conditional probability Prij h | u(f)based on the values of precursors is the same for all cells, and the space indices of
probability Pr may be dropped. This condition is widely accepted in constructions of
the forecasting algoritms but rarely formulated explicitly. However, the probability of
a seismic event depends to a large extent on the local properties of the area. Hence,
the quality of the forecasting depends on how adequately these properties are sum-
marized in the precursors. This formalism properly describes the space inhomogenuity
of the physical space because the stationary joint distribution of Prij h,f≤ x for an
arbitrary precursor f depends, in general, on the position of the cell in the domain A.
Below we will use the distributions of precursors and indicators of seismic events in
domain A that do not depend on the spatial coordinates and have the following form
PrA h,f≤ x =1
I · J∑
(i,j )∈A
Prij h,f≤ x ,
PA(x) ≡PrA f≤ x =1
I · J∑
(i,j )∈A
Prij f ≤ x ,
6 V.Gertsik, M.Kelbert, A.Krichevets
pA ≡PrA h= 1 =1
I · J∑
(i,j )∈A
Prij h = 1 ,
(iii-d) Note that assumption (iii) implies that the conditional probabilities Pr (h|u(f))are computed via the probabilities PrA h,f≤ x only.
The properties listed above are sufficient to obtain the point estimates for the
conditional probabilities of seismic events under conditions formulated in terms of the
values of precursors. However, additional assumption are required for a testing of the
forecasting algorithm:
(iv) The random variables f(i, j, k), are conditionally independent under condition that
h(i, j, k) = 1.Again, these conditions are not exactly true, however they may be treated as a
reasonable approximation to reality. Indeed, if the threshold M0 is sufficiently high
than the strong earthquakes may be treated as rare events, and the cells where they
are observed are far apart with a high probability. Any two events related to cells
separated by the time intervals ∆t are asymptotically independent as ∆t → ∞ because
the seismic process has decaying correlations (the mixing property in the language of
random processes). The loss of dependence (or decaying memory) is related to the
physical phemonema such as healing of the defects in the rocks, relaxation of strength
due to viscosity, etc. As usual in physical theories, we accept an idealized model of the
real phenomena applying this asymptotic property for large but finite intervals between
localizations of seismic events.
The independence of strong earthquakes is not a new assumption, in the case of con-
tinuous space-time it is equivalent to the assumption that the locations of these events
form a Poisson random field.. (Note that the distribution of strong earthquake should
be homogeneous in space, because there is no information about the heterogeneity a
priori .) The Poisson hypothesis is used in many papers, see, e.g. [Harte and Vere-Jones
, 2005]. It is very natural for the analysis of the «tails» of the Gutenberg-Richter law
for large magnitudes [Pisarenko et al., 2008]. Summing up, the development of the
strict mathematical theory of earthquake forecasting does not require any additional
assumption except those routinely accepted in the existing algorithms but usually not
formulated explicitly.
4 The standard form for precursors
The correct solution of the forecasting problem given the values of precursors f(i, j, k) =(
f1(i, j, k), . . . , fQ(i, j, k))
is provided by the estimate of conditional probability Prh | f(i , j , k)of the indicator of seismic event in the cell (i, j, k). In practice this solution may be
difficult to obtain because the number of events in catalog is not sufficient.
Indeed, the range of value of a scalar precursor is usually divided into a number
M of intervals, and only a few events are registered for any such interval. For a Q-
dimensional precursor the number of Q-dimensional rectangles, covering the range, is
already MQ, and majority of them contains 0 event. Only a small number of such
rectangles contains one or more events, that is the precision of such an estimate of
conditional probability is usually too low to have any practical value.
For this reason one constructs a new scalar precursor in the form of the scalar
function of component of the vector precursor, and optimize its predictive power. This
Earthquake forecasting: Statistics and Information 7
approach leads to additional complication as the units of measurement and the phys-
ical sense of different components of precursor are substantially different. In order to
overcome this problem one uses some transformation to reduce all the components of
the precursor to a standard form with the same sense and range of values.
Let us transform all the precursors fq(i, j, k), q = 1, ..., Q to variables with the
values in [0,1] providing estimates of conditional probabilities. So, after some trans-
formation F we obtain an estimate of Pr h = 1 | u(f(i, j, k)) = 1, where u is a char-
acteristic function of some interval B, i.e., the probability of event h(i, j, k) = 1 under
condition that this precursor takes the value f(i, j, k) ∈ B.
The transformation F of a scalar precursor f(i, j, k) is defined as follows. Fix an
arbitrary small number ε. Let L be a number of cells (i, j, k) such that h(i, j, k) = 1,and Zl, l = 1, . . . , L, be the ordered statistics, i.e., the values f(i, j, k) in these cells
listed in non-decreasing order. Define a new sequence zm, m = 0, ...,M, from the
ordered statistics Zl by the following recursion: z0 = −∞, zm is defined as the first
point in the sequence Zl, such that zm − zm−1 ≥ ε. Next, construct the sequence
z∗m = zm + (zm+1 − zm)/2,m = 1, . . . ,M − 1, and add the auxiliary elements z∗0 =−∞, z∗M = ∞. Define also a sequence nm, m = 1, . . . ,M , where nm equals to the
number of values in the sequence Zl, such that z∗m−1 ≤ Zl < z∗m. Finally, define the
numbers Nm, m = 1, ...,M counting all cells such that z∗m−1 ≤ f(i, j, k) < z∗m, m =
1, ...M . Observe that∑M
m=1 nm = L,∑M
m=1 Nm = N , and use the ratios
λ =L
N(1)
as estimate of unconditional probability of a seismic event in a given cell
pA ≡ PrA h(i , j , k) = 1 =
ˆ ∞
−∞Pr h = 1 | x dPA(x). (2)
The transformation F is defined as follows
g = Ff(i, j, k) =nm
Nm, if z∗m−1 ≤ f(i, j, k) < z∗m, m = 1, . . . ,M. (3)
This definition implies that transformation F replace the value of precursor for
the frequency, i.e., the ratio of a number of cells containing a seismic event and the
values of precursor from [z∗m−1, z∗m) to the number of cells with the value of precursor
in this range. These frequencies are the natural estimates of conditional probabili-
ties Pr
h = 1 |z∗m−1 ≤ f < z∗m
, m = 1, . . . ,M , computed with respect to stationary
distribution PA(x):
Pr
h = 1 |z∗m−1 ≤ f < z∗m
=
´ z∗
m
z∗
m−1
Pr h = 1 | x dPΩ(x)´ z∗
m
z∗
m−1
dPA(x). (4)
(This conditional probability can be written as Pr h = 1 |u(f ) , where u is the char-
acteristic function of interval [z∗m−1, z∗m)). The function g has a stepwise shape, and
the length of the step in bounded from below by ε. It can be checked that there exist
the limit g = limε→0
limK→∞
g = Pr h = 1 |f .The estimates of conditional probabilities in terms of the function g are quite rough
because typically the numbers nm are of the order 1. As a final result we will present
below more sharp but less detailed estimates of conditional probabilities and confidence
intervals for them.
8 V.Gertsik, M.Kelbert, A.Krichevets
5 Combinations of precursors
There are many ways to construct a single scalar precursor based on the vector pre-
cursor (Ffq, q = 1, . . . , Q). Each such construction inevitably contains a number of
parameters or degrees of freedom. These parameters (including the parameters used
for construction of the precursors themselves) should be selected in a way to optimize
the predictive power of the forecasting algorithm. The optimization procedure will be
presented below, its goal is to adapt the parameters of precursors to a given catalog
of earthquakes, that is to obtain the best possible retrospective forecast. However, this
adaptation procedure creates a "ghost" information related with the specific features
of the given catalog but not present in physical propertities of real seismicity. This
"ghost" information will not be reproduced if the algorithm is applied to another cat-
alog of earthquakes. It is necessary to increase the volume of the catalog and to reduce
the number of free parameters to get rid of this "ghost" information.. Clearly, the first
goal requires the considerable increase of the observation period and may be achieved
in the remote future only. So, one concentrates on the reduction of number of degrees
of freedom. The simplest ansatz including Q− 1 parameters is the linear combination
f∗ = Ff1 +
Q∑
q=2
cq−1F fq . (5)
As a strictly monotonic function of precursor is a precursor itself the log-linear combi-
nation is an equally suitable choice
f∗ = ln (Ff1) +
Q∑
q=2
cq−1 ln (F fq) , (6)
Here cq , q = 1, ..., Q− 1 are free parameters. The result of the procedure has the form
g = Ff∗.
6 Alarm levels, point and interval estimations
In view of (3) the precursor g is the set of estimates for probabilities
Pr
h = 1 |z∗l−1 ≤ f (i , j , k) < z∗l
, l = 1, ..., L(f).
Its serious drawback is that typically
z∗l−1 ≤ f(i, j, k) < z∗l
correspond to single
events, and therefore the precision of these estimates is very low (the confidence in-
tervals discussed below may be taken as a convenient measure of precision). In order
to increase the precision it is recommended to use the larger cells containing a larger
number of events, that is a more coarse covering of the space where the precursor takes
its values. In a sense, the precision of the estimation and the localization of the pre-
cursor values in its time-space region are related by a kind of "uncertainty principle":
the more precise estimate one wants to get the more coarse is the time-space range of
their values and vice versa.
We adapt the following approach in order to achieve a reasonable compromise.
1. For fixed thresholds as, s = 1, ..., S + 1, a1 = 1, as < as+1, aS+1 = 0, we
define S possible alarm levels as+1 ≤ g(i, j, k) < as and subsets Ωs, s = 1, ..., S, of
Earthquake forecasting: Statistics and Information 9
the set ΩK corresponding to alarm levels, i.e., Ωs is a set of cells of ΩK , such that
as+1 ≤ g(i, j, k) < asThere are different ways to choose the number S of alarm levels and the thresholds
as, s = 2, ..., S. Say, fix S = 5, and select as = 10−α(s−1). This is a natural choice of
the alarm level because at α = 1 it corresponds to decimal places of the estimate of the
conditional probability given by the precursor. The problem with S = 2, i.e., two-level
alarm, may be reduced to the hypothesis testing and discussed in more details below.
2. Compute the point estimates θs of probabilities Pr h = 1 |as+1 ≤ g(i , j , k) < as ,s = 1, ..., S, obtained via the distribution PΩ(x) of precursor g in the same way as in
(4). The property (iv) implies that for any domain Ωs the binary random variables
h(i, j, k) are independent and identically distributed, i.e
Pr h = 1 |as+1 ≤ g(i , j , k) < as ≡ ps,
and the unbiased estimate of ps takes the form
θs =ms
ns(7)
where ns stands for the number of cells in domain Ωs, and by ms we denote the number
of cells in Ωs containing seismic events.
3. The random variable ms takes integer values between 0 and ns. The probabili-
ties of these values are computed via the well-known Bernoulli formula Pr(ms = k) =(ns
k ) pks(1 − ps)
ns−k. Let us specify the confidence interval covering the unknown pa-
rameter ps with the confidence level γ. In view of the integral Mouvre-Laplace theorem
for ns large enough the statistics(θs−ps)
√ns√
ps(1−ps)is approximately Gaussian N(0,1) with
zero mean and unit variance. Note that the values ns increase with time. Omitting
straightforward calculations and replacing the parameter ps by its estimate θs we ob-
tain that θ−s < θs<θ+s , where θ−s = θs − tγ√
θs(1−θs)√ns
, θ+s = θs +tγ√
θs(1−θs)√ns
, and
tγ is the solution of equation Φ(tγ) = γ2 . Here Φ stands for the standard Gaussian
distribution function.
4. As a result of these considerations we introduce ’the precursor of alarms’ which
indicates the alarm level: R(f(i, j, k)) = s(i, j, k). It will be used for calculations of
point estimate and the confidence inteval in the form θ−s(i,j,k)
< θs(i,j,k) < θ+s(i,j,k)
.This result will be use for prospective forecasting procedure.
7 The information gain and the precursor quality
The construction of a ’combined’ precursor R involves parameters from formula (5) or
(6) as well as parameters which appear in definition of each individual precursors fq . It
is natural to optimize the forecasting algorithm in such a way that the information gain
related to the seismic events is maximal. In one-dimensional case the information gain
as a measure of the forecast efficiency was first intoduced by Vere-Jones [Vere-Jones,
1998]. Here we exploit his ideas in the case of multidimensional space-time process.
Remind the notions of the entropy and information. Putting aside the mathematical
subtlety (see [Kelbert, Suhov, 2013] for details) we follow below an intuitive approach
of the book [Prohorov, Rozanov, 1969]. The information containing in a given text is,
basically, the length of the shortest compression of this text without the loss of its
content. The smallest length S of the sequence of digits 0 and 1 (in a binary code) for
10 V.Gertsik, M.Kelbert, A.Krichevets
counting N different objects satisfies the relations 0 ≤ S− log2 N ≤ 1. So, the quantity
S ≈ log2 N characterizes the shortest length of coding the numbers of N objects.
Consider an experiment that can produce one of N non-intersecting events А1, . . . ,АN
with probabilities q1, . . . , qN , respectively, q1 + . . . + qN = 1. A message informing
about the outcomes of n such independent identical experiments may look as a se-
quence (Ai1 , . . . , Ain), where Aik is the outcome of the experiment k. But for long
enough series of observations the frequency ni/n of event Аi is very close to its prob-
ability qi. It means that in our message (Ai1 , . . ., Ain) the event Аi appears ni times.
The number of such messages is
Nn =n!
n1!...nN !.
By the Stirling formula the length of the shortest coding of these messages
Sn ≈ log2 Nn ≈ −n
N∑
i=1
qi log2 qi.
The quantity Sn measures the uncertainty of the given experiment before its start,
in our case we are looking for one of possible outcomes of n independent trials. The
specific measure of uncertainty for one trial
1
nSn =
1
nSn(q1, . . . , qN ) = −
N∑
i=1
qi log2 qi
is known as Shannon’s entropy of distribution q1, . . . , qN (in physical literature it is
also known as a measure of chaos or disorder). After one trial the uncertainty about
the future outcomes decreases by the value S = Sn − Sn−1, this decrement equals to
the information gain I = S, obtained as a result of single trial.
The quantity
S(h) = −pA log2 pA − (1− pA) log2(1− pA) (8)
is the (unconditional) entropy of distribution for indicator of seismic event h in a space-
time cell in the absence of any precursors. The conditional entropy S(h | as+1 ≤ g <
as) under condition that in the cell (i, j, k) the alarm level s is set up equals
S(h | as+1 ≤ g < as) = −ps log2 ps − (1− ps) log2(1− ps)
The expected conditional entropy SR(h) of indicator of seismic events where the aver-
aging in taken by the distribution of precursors R takes the form
SR(h) = −S∑
s=1
[ps log2 ps + (1− ps) log2(1− ps)]PA(as+1 ≤ g < as) (9)
We conclude that the knowledge of the precursor values helps to reduce the un-
certainty about the future experiment by S(h)− SR(h) which is precisely information
I(R, h) obtained from the precursor. Taking into account (8), (9) and the fact that
pA =S∑
s=1
psPA(as+1 ≤ g(i, j, k) < as)
Earthquake forecasting: Statistics and Information 11
we specify the information gain as
I(R, h) =S∑
s=1
[
ps log2pspA
+ (1− ps) log21− ps1− pA
]
PA(as+1 ≤ g < as).
By analogy with the one-dimensional case [Kolmogorov, 1965] the quantity I(R, h)may the called the mutual information about the random field h that may be obtained
from observations of random field R. It is known that the information I(R, h) is non-
negative and equals to 0 if and only if the random fields h and R are independent.
This mutual information I(R, h) takes its maximal value S(h) in an idealized case
of absolutely exact forecast. The mutual information quantifies the information that
the distributions of precursors contribute to that of the indicator of seismic event. For
this reason it may be considered as an adequate scalar estimate for the quality of the
forecast.
The quantity I(R, h) depends on the cell size, i.e., on the space discretization
length a and time interval ∆t. We need a formal test to compare precursors defined
for different size of the discretization cells. For this aim let us introduce the so-called
’efficiency’ of precursors as the ratio of information gains
r(R, h) =I(R, h)
S(h).
This efficiency varies between 0 and 1 and serves as a natural estimate of information
quality of precursors. It allows to compare different forecasting algorithms and select
the best one.
A natural estimate of S(h) based on (1) and (2) is defined as follows
S(h) = −λ log2 λ− (1− λ) log2(1− λ). (10)
Taking into account (7) and using an estimate of PA(as+1 ≤ g < as) in the form of
ratio τs = ns
N , we construct an estimate of I(R, h) as follows
I(R, h) =S∑
s=1
[
θs log2θsλ
+ (1− θs) log21− θs1− λ
]
τs, (11)
r(R, h) =I(R, h)
S(h). (12)
Remark 2 The economical quality of forecast. A natural economic measure for a quality
of binary forecast is the economic risk or damage r related to the earthquakes and the
necessary protective measures. In mathematical statistics the risk is defined as the
expectation of the loss function, in our case there are two types of losses: damage and
expenses related to protection. For each cell of our grid the risk may be specified by
the formula
r = αPrh(i , j , k) = 1, η(i , j , k) = 0+ βPrh(i , j , k) = 0, η(i , j , k) = 1++γPrh(i , j , k) = 1, η(i , j , k) = 1,
here α stands for the average damage from a seismic event; β stands for the average
expenses for protection after a seismic alarm is announced; γ stands for the average
12 V.Gertsik, M.Kelbert, A.Krichevets
damage after the alarm, γ = α+β− δ, where δ is the damage prevented by the alarm.
The coefficient in front of Prh(i , j , k) = 0, η(i , j , k) = 0, obviously, equals 0, because
in the absence both of a seismic event and an alarm there is no loss of any kind. Clearly,
only the case when δ > β is economically justified, i.e., the gain from the prevention
measures is positive. Obviously, δ should be less than α+β, i.e., an earthquake cannot
be profitable. Taking into account that α, β and γ depend on the geographical position
of the cell, we write the total risk as the summation over all cells in the region of a
given forecast. In the simplest case of the absence of the spacial component, when a
single cell represents a region of forecast, the expression for the risk is simplified as
follows r = αλν + βτ + γλ(1− ν).
However, the risk r, which is very useful for economical considerations and as a basis
for an administrative decision, could hardly be used as a criteria for quality of seismic
prediction. First of all, it cannot be computed in a consistent way because the coeffi-
cients α, β and γ are not known in practice, and hence no effective way of its numerical
evaluation is known. The computation of these coefficients is a difficult economic prob-
lem and goes far beyond of the competence of geophysicists. On the other hand, the
readiness of the authority to commit resources to solving the problem depends on the
quality of the geophysical forecast. This situation leads to a vicious cirle.
The second drawback of the economic risk as a criterion for the quality of predic-
tion is related to the fact that it depends on many factors which have no relation to
geophysics or earthquake prediction. These factors inlude the density of population,
the number and size of industrial enterprises, infrastructure, etc. It also depends on
subjective factors such as the williness of authorities to use resources for prevention of
the damage from earthquakes. The natural sciences could hardly accept the criteria for
the forecast quality which depend on the type of state organization, priorities of ruling
parties, results of the recent elections, etc.
Remark 3 It seems reasonable to introduce a penalty related to the number of super-
fluous parameters in evaluating the quality of forecast pointing to the natural analogy
with the Akaike test [Akaike, 1974] and similar methods in information theory. In our
context the main parameter of importance is r(R, t) and its limit as t → ∞. This
quantity does not involve the number of parameters directly. Probably, the rate of
convergence depends on the number of parameters but this dependence is not studied
yet.
8 The forecasting procedure
The number of time intervals, i.e., the number of observation N used in the construc-
tion of estimates increases with the growth of observation time. So, the computation
procedure requires constant innovations. On the other hand some computation time
is required to ’adapt’ the model parameters to the updated information about seismic
events via an iterative procedure. For these reasons we propose the following forecasting
algorithm.
1. Given initial parameter values at the moment tK−1 = t0 + (K − 2)∆t we
optimize them to obtain the maximum of efficiency r(R, h) of precursor in domain
ΩK−1. For this aim the Monte-Carlo methods is helpful: one perturbs the current
values of parameters randomly and adapts the new values if the efficiency increases.
The process continues before the value of efficiency stabilized, this may give a local
Earthquake forecasting: Statistics and Information 13
maximum, so the precedure is repeated sufficient number of times. The choice of initial
value on the first step of optimization procedure is somewhat arbitrary but a reasonable
iteration procedure usually leads to consistent results. The opmization procedure takes
the period of time tK−1 < t ≤ tK .
2. Next, we construct the forecast in the following way. At the moment tK the
values of precursor g in each cell (i, j, K + 1) is computed with optimized parameters.
Based on these parameter values the alarm levels, the point estimates and confidence
intervals are computed in each cell as well as the values of efficiency of precursors.
3. The estimates of stationary probabilities of seismic events in the cell θ(i, j) are
defined as follows:
θ(i, j) =1
K
K∑
k=1
θs(i,j,k).
they can be used for creation of the of the variant of the maps of seismic hazard in the
region.
9 Retrospective and prospective informativities
The efficiency of precursor which is achieved as a result of parameters optimization
could be considered as retrospective as it is constructed by the precursors adaptation
to the historical catalogs of seismic events. The prospective efficiency for the space-time
domain Ω∗ containing the cell in the ’future’ is based on the forecast. It is computed
via formulas (10), (11), (12) with the only difference that domain Ω∗s consists from
the cells where the forecasted alarm level is s. The efficiency of prospective forcast
is smaller compared with the retrospective efficiency, however approaches this value
with time. In principle, the prospective efficiency is an ultimate criteria of precursors
quality and the retrospective efficiency could serve only for the preliminary selection
of precursors and their adaptation to the past history of seismic events.
10 Testing of the forecasting algorithm
The efficiency of precursor could be computed exactly only in an idealized case of infi-
nite observation time. However, its estimate may be obtained based on the observation
over a finite time interval. So, if an estimate produces a non-zero value not necessarily
the real effects is present. It may be simply a random fluctuation even if the precursor
provides no information about the future earthquake. For this reason we would like to
check the hypothesis H0 about the independence of a precursor and an event indica-
tor with a reasonable level of confidence. In case the hypothesis is rejected one have
additional assurance that the forecasting is real, not just a "ghost".
So, consider the distributions
PA(x) =1
I · J∑
(i,j )∈Ω
Prij g(i , j , k) ≤ x ,
and
P′A(x) =
1
I · J∑
(i,j )∈A
Prij g(i , j , k) ≤ x |h(i, j, k) = 1
14 V.Gertsik, M.Kelbert, A.Krichevets
The function PA(P−1A
(y)) = y of variable y = PA(x) provides an uniform distribution
F ∗(y) = Pr(
ξ ≤ y)
of some random variable ξ on [0,1]. Next, consider a distribution
function G(y) = P ′A(P
−1A
(y)) on [0,1], and use a parametric representation for abcissa
PA(x) and ordinate P ′A(x). If random fields g and h are independent the distribution
functions PA(x) = P ′A(x) and G(y) are uniform. So, the hypothesis about the absence
of forecasting, i.e., about the independence of g and h, is equivalent to the hypothesis
H0 that the distribution G(y) is uniform.
The empirical distribution GL(y) related to G(y) is defined as follows. Denote by
ul, l = 1, ...L the values of the function g(i, j, k) sorted in the non-decreasing order
and beloning to the cells where h(i, j, k) = 1. Let nl be the numbers of cells such that
h(i, j, k) = 1, g(i, j, k) = ul. Denote by m(ul) the numbers of cells from Ω such that
g(i, j, k) < ul, and define the empirical distribution GL(y) as a step-wise function with
GL(0) = 0 and positive jumps of the size nl
L at points yl =m(ul)N .
The well-known methods of hypothesis testing requires that the function GL(y)has the same shape as for independent trials, i.e., random variables ul, l = 1, ...L are
independent in view of axiom (iv). Naturally, we accept the precursors such that the
hypothesis H0 is rejected with the reasonable level of confidence. (Remind, that the
hypothesis is accepted if and only if its logical negation could be rejected based on the
available observations. The fact that the hypothesis cannot be rejected does not mean
at all that it should be accepted, it only means that the available observations don’t
contradict this hypothesis. Say, the well-known fact that "The Sun rise in the East"
does not contradict to our hypothesis, however it may not be considered as a ground
for its acceptance.) For large values of L the Kolmogorov statistics [Kolmogorov, 1933a]
is helpful for this aim
DL = sup | GL(y)− y |
with an asymptotic distribution
limL→∞
Pr√
LDL ≤ z
=∞∑
k=−∞
(−1)k e−2k2 z2
, z > 0,
or Smirnov’s statistics [Smirnov, 1939]
D+L = sup [GL(y)− y] ,
D−L = − inf [GL(y)− y] ,
with asymptotic distribution
limL→∞
Pr√
LD+L ≤ z
= limL→∞
Pr√
LD−L ≤ z
= 1− e−2z2
, z > 0.
The asymptotic expressions for these statistics can be used for L > 20 ([Bolshev,
Smirnov, 1965])..
Earthquake forecasting: Statistics and Information 15
11 The binary alarm and the hypothesis testing
The prediction is the form of forecast when an alarm is announced in a given cell
without a preliminary evaluation of probability of seismic event. In this case we can
estimate the probabilities of events too. (If the alarm is announced in an arbitrary
domain Ω we set up an alarm if at least haph of the cell of our model is occupied by
alarm.).
Let M be the number of cells in Ω which are in the state of alarm, M0 be the
number of cells where the seismic event is present but no alarm was announced (the
number of ’missed targets’). Denote by τ = MN the share of the cells with alarm
announced, λ = LN the share of the cells with seismic events, and ν = M0
M the share
of missed targets. Let a random variable η(i, j, k), equal 1 if an alarm is announced
in the cell (i, j, k), and 0 otherwise. Obviously, the estimate of conditional probability
Pr h(i , j , k) = 1 | η(i , j , k) = 1 of the seismic event under the condition of alarm isλ(1−ν)
τ , and the estimate of conditional probability Pr h(i , j , k) = 1 | η(i , j , k) = 0of the seismic event under the condition of no alarm is λν
1−τ .
If the alarm is announced according to the procedure described in Section 5 the
threshold a1 specifying the acceptable domain of values for g(i, j, k) should be treated
as a free parameter and selected by maximizing the information efficiency r(η, h). The
estimate of information increase for given values of τ and ν equals
I(η, h) = λ(1− ν) log21− ν
τ+ λν log2
ν
1− τ+
+ [τ − λ(1− ν)] log2τ − λ(1− ν)
(1− λ)τ+ (1− τ − λν) log2
1− τ − λν
(1− λ)(1− τ).
The value of η(i, j, k) characterizes the results of checking two mutually exclusive
simple hypothesis:
H0: the distribution of g(i, j, k) has the form P0A(x) ≡ P rA g(i , j , k) ≤ x | h(i ,j ,k) = 0,
implying ’no seismic events’,
or
H1 : the distribution of g(i, j, k) has the form P1A(x) ≡ P rA g(i , j , k)≤ x | h(i ,j ,k) = 1,
implying the presence of seismic event.
Statistics for checking of these hypothesis is the precursor g(i, j, k), and the critical
domain for H0 has the form g(i, j, k) ≥ a1. (If usual method of alarm announcement
is used the relevant precursor plays the rôle of statistics and the critical domain is
defined by the rule of the alarm announcement). The probability of first type error
α = Pr η(i , j , k) = 1 | h(i , j , k) = 0 ,
it is estimated as τ−λ(1−ν)1−λ
. The probability of second type error
β = Pr η(i , j , k) = 0 | h(i , j , k) = 1 ,
it is estimated as ν. (Note that due to condition (iii) any test used for the checking
these hypothesis should not depend on the coordinates of the cell).
The Neyman-Pearson theory allows to define the domain of images of all possible
criteriaall possible criteria: in coordinates (α, β) it is a convex domain with a boundary
16 V.Gertsik, M.Kelbert, A.Krichevets
Γ which corresponds to the set of uniformly most powerful tests. This family may
be defined in terms of the likelihood ratio Λ(x) =p1
A(x)
p0
A(x)under condition that the
distributions P1A(x) and P0
A(x) has densities p1A(x) and p0
A(x):
η(i, j, k) = 1 if Λ(x) > ω,
η(i, j, k) = 0 if Λ(x) < ω
where ω denotes the threshold. In the paper [Gercsik, 2004] we demonstrated that
among all the tests with the images on the boundary Γ there exists three different best
tests. Here the term "best" may be understood in three different sense, i.e., maximizing
the variational, correlational and informational efficiency. The most relevant criteria is
the informational efficiency r(η, h).The well-known Molchan’s error diagram [Molchan, 1990] where the probability of
the first kind error is estimated by τ is constructed in the same way. However, it involve
a comparison of two intersecting hypothesis:
H0: the distribution g(i, j, k) has the form P0A(x) ≡ PrA g(i , j , k)≤ x, i.e., the
seismic event could "either happen or not happen", and
H1 : the distribution g(i, j, k) has the form P1A(x) ≡ PrA g(i , j , k) ≤ x | h(i ,j ,k) = 1,
i.e., the seismic event "will happen"
.Note that the rejection of hypothesis H0 leads to absurd results.
Remark 4 In the paper [Molchan and Keilis-Borok, 2008] the area of the alarm domain
is defined in terms of non-homogeneous measure depending on the spacial coordinates,
in terms of our paper it may be denoted as θ(i, j). i.e., τ ∽
∑
i,j θ(i, j)η(i, j, k). This
approach is used to eliminate the decrease of the share of alarmed sites τ with the
extension of the domain when a purely safe and aseismic territory is included into
consideration. It would be well-justified if the quantity τ could be accepted as an ade-
quate criterion of the quality of forecast in its own right. On the other hand, it can be
demonstrated that the information efficiency r(η, h) converges to a non-zero value 1−ν
when the number of cells with an alarm is fixed but the total number of cells tends
to infinity. An inhomogeneous area of the territory under forecast which is propor-
tional to θ(i, j) does not enable us to calculate the informational efficiency. Moreover,
it possesses a number of unnatural features from the point of view of evaluation of
economical damage. A seismic event in the territory of low seismicity is more costly
because no precautions are taken to prevent the damage of infrastructure. However,
in this inhomogeneous area an alarm announced in an aseismic territory will have a
smaller contribution than an alarm in a seismically active territory where the losses
would be in fact smaller. We conclude that this approach ’hides’ the most costly events
and does not provide a reasonable estimate of economic damage.
12 The choice of precursors
We use the term ’empiric precursor of earthquake’ for any observable characterisric de-
rived from the catalog only which provides for this catalog a reasonable retrospective
forcast of seismic events and not derived from basic physical conception of seismicity
(say, the periods of relatice calm, deviation of some basic characteristic from a long-time
average , etc). In contrast, the physical precursors are de- rived from some of physical
processe and characterize physical quantities (stress fields, strength, concentration of
Earthquake forecasting: Statistics and Information 17
cracks, etc.) or well-defined physical processes (i.e., phase transitions, cracks propaga-
tions, etc.) In the meteorological forecast the danger of using empirical precursors was
highlighted by A. Kolmogorov in 1933 [Kolmogorov, 1933]. From that time the mete-
orological forecast relies on the physical precursors which are theoretically justified by
the models of atmospheric dynamics. Below we will present A. Kolmogorov’s argument
adapted to the case of seismic forecast. This demonstrates that the purely empirical
precursors work well only for the given catalog from which they are derived. However,
their eficiency deteriorates drastically when they are applied to any other independent
catalog.
Consider a group of k empirical precursors used for a forecast and and selected from
a set of n such groups. According to A. Kolmogorov’s remark the number k is typically
rather small. This is related to the fact that a number of strong earthquakes in catalog
is unlikely to exceed a few dozen. As the values of precursors are random there exists
a small probability p that the efficiency of the forecast exceeds the given threshold
С. Then the probability of event r(R, h)≤С equals 1–p, and the probability of event
r(R, h) > С for at least one collection of precursors equals P = 1− (1− p)n and tends
to 1 as n → ∞.. (According to Kolmogorov some arbitrariness of the assumption of
independence is compensated by the large number of collections.)
Summing up, if the number of groups is large enough with probability close to
1 it is possible to find a group giving an effective retrospective forecast for a given
catalog. In practice this is always the case as the number of empirical precursors could
be increased indefinitely by variation of real parameters used in their construction. It is
important to note that for such a group, which is highly eficient for the initial catalog,
the probability that the eficiency is greater than C is still equal to p for any other
catalog. In other words the larger the number of the groups of empirical precursors the
less reliable forecast is. So, the collection of a large list of the empirical precursors is
counter-productive.
Much more reliable are the physical precursors intrinsicly connected with the phys-
ical processes which preserve their values with the change of sample. The probability
to find such a set of precursors by pure empirical choice is negligible because they are
very rare in the immense collection of all possible precursors.
13 Image identification
The possibility to use the pattern recognition formalism in seismic forecast is totally
based on the acceptance of deterministic model of seismicity. It is necessary to assume
that in principle there exits such a group of precursors which allows to determine with
certainly whether a strong earthquake will happen or not. In this case one believes that
all random errors are related to the incompleteness of this set of precursors.But if the
seismicity is a random process then the image appears only after the earthquake and
before it any set of values for precursors cannot guarantee the possible outcome and
only the relevant probabilities may be a subject of scientific study. After the discovery
of dynamic instability and generators of stochastic behavior of dynamical systems the
deterministic model of seismicity is cast in doubts. Its potential acceptance requires
substantial evidence which hardly exist at present.
In any case the results of pattern recognition procedure (i.e., a binary alarm) are
useful if they are considered alongside with the results of statistical tests. They allows
to calculate the estimate of probabilities of seismic events and informational efficiency.
18 V.Gertsik, M.Kelbert, A.Krichevets
However, the section of ’features’ for pattern recognition leads to the same difficul-
ties as the selection of precursors: the ’features’ based on the observations only and not
related to the physics of earthquakes are not helpful, and any hopes for ’perceptron
education’ are not grounded. A successful supervised recognition is possible if the fea-
tures has proved causal relation with pattern. This principle is illustrated by a simple
but important theorem by A.N. Krichevets.
Theorem 1 Let A be a finite set, B1, B2 ⊂ A,B1 ∩B2 = ∅. We say that B1 and B2
are finite educational samples. Let X ∈ A, X /∈ B1 ∪B2 be a new object. Then among
all classifications, i.e., subsets (A1, A2) such that B1 ⊂ A1, B2 ⊂ A2, A1 ∪ A2 = A,
A1 ∩ A2 = ∅ satisfying condition that either B1 ∪X ⊂ A1 or B2 ∪ X ⊂ A2 exactly a
half classifies X as an object of sample B1 and a half classifies X as an object from
B2.
Proof It is easy to define a one-to-one between classifications. Indeed, if A1, A2 ,A1 ∪ A2 = A, is a classification such that B1 ⊂ A1, X ⊂ A1, B2 ⊂ A2, one maps it
into the unique classification
A′1, A
′2
such that B1 ⊂ A′1, X ⊂ A′
2, B2 ⊂ A′2, where
A′1 = A1 \X, A′
2 = A1 ∪X.
Corollary 1 A supervised pattern recognition is impossible. After the leaning procedure
the probability to classify correctly a new object is the same as before leaning, i.e., 1/2.
14 Demonstration of algorithm
A preliminary version of the forecast algorithm described above was used in the paper
[Ghertzik, 2008] for California and the Sumatra-Andaman earthquake region. These
computations serve as a demonstration of the efficiency of the method but their actual
results should be taken with a pinch of salt because the selection of precursors does not
appear well-justified from the modern point of view: the number of free parameters to
be adapted in the precursor ‘"stress indicator" is too large.
Califormia region. The catalog Global Hypocenter Data Base CD-ROM NEIC/USGS,
Denver, CO, 1989, together with data from the site NEIC/USGS PDE (ftp://hazard.cr.usgs.gov)
for earthquakes with magnitudes M ≥ 4.0 with epicenters between 113−129 of west-
ern longitude and 31 − 43 of northern latitude was used for parameter adaptation.
The initial time t0 was selected by subtracting from the time of actual computation,
08.03.2006, an integer number of half-year intervals such that t0 fits the first half of
the year 1936. (The final time 08.03.2006 could be considered as an initial moment for
constructing half-year forecast forward up to the date of the latest earthquake available
in the catalog). During the computation the time interval from the first half of 1936 to
the first half of 1976 was used for relaxation of the zero initial data used for precursors.
After this date the catalog for the earthquakes with magnitude M ≥ 4.0 was used to
estimate the probabilities of strong earthquakes with magnitude M ≥ 6.0 up to the
moment of actual forecast. Note that the adapted restriction to include into consider-
ation only earthquakes with magnitudes M ≥ 4.0 is a severe restriction. It decreases
the precision of precursor computation and therefore, if a prediction is successful, in-
creases the degree of confidence to the predictor choice. We choose a = 150km as a size
of the spacial lattice, and ∆t = 6 months as a time-step. Retrospective forecast was
performed with 5 alarm levels defined by the thresholds as = 10−α(5−s), s = 1, . . . , 4and α = 0.75. (Due to too short time step no alarms was registered on the lowest level
Earthquake forecasting: Statistics and Information 19
when parameter α = 1 was selected). In order to reduce the influence of the boundary
conditions the large square covering all the seismic events in the catalog used in the
computations was reduced by two layers of elementary cells from each boundary. As
a result of optimization the forecast information efficiency of 0.526 was achieved, i.e.,
the forecasting algorithm applied to the given catalog extracted from it about 53% of
all available information about seismic events. It seems that this result could be only
partially explained by a lucky selection of precursors: another contributor to the high
efficiency of the algorithm is the adaptation of the parameters to the features of the
specific catalog. The influence of this artificial information may be reduced only with
the increase of the observation interval.
Accepting the rule of binary alarm announcement in the cells from group 1 and 2
from 5 levels possible one obtains that the space-time share of alarmed cells is 3.4%and the share of missed targets is 18.2%. This result is comparable with the best
forecasts available in the literature and obtained by other methods (in the cases when
the quantitive parameters of algorithms are presented in the publications). When the
forecast was constructed in the future we obtained that the estimate of probability of
a strong earthquake anywhere in the area under study is 0.174, and the maximal point
estimate of an event in any individual cell is 0.071. As a whole the seismic situation
in California did not look too alarming. Indeed, there were no strong earthquakes in
California in the next half-year.
SAE region. We have conducted a retrospective forecast of strong earthquakes
with magnitudes M ≥ 7.0 for the whole region where the Sumatra-Andoman earth-
quake (SAE) happened on 26.12.2004 with magnitude M = 9.3. The catalog Global
Hypocenter Data Base CD-ROM NEIC/USGS, Denver, CO, 1989, together with the
data from the cite NEIC/USGS PDE (ftp://hazard.cr.usgs.gov) for earthquakes with
magnitudes M ≥ 5.5 with epicenters between 84.3 − 128 of eastern longitude and
20 − 26 of northern latitude was used for the parameters adaptation. The initial
moment of time t0 was selected by subtracting from the time of actual computation,
10.11.2004, an integer number of half-year intervals such that t0 fits the first half of the
year 1936. (The final time was selected in such a way that the next half-year period
covers SAE and its powerful aftershock). During the computation the time interval
from the first half of 1936 to the first half of 1976 was used for relaxation of the zero
initial data used for precursors. After this date the catalog was used to estimate the
probabilities of strong earthquakes with magnitude M ≥ 7.0 up to the moment of ac-
tual forecast. (For magnitude M ≥ 7.5 the number of seismic events was not sufficient
for reliable forecast because the 5%-confidence intervals strongly overlapped). In this
case the restriction to include into consideration only earthquakes with magnitudes
M ≥ 5.5 was adapted. As before, it decreases the precision of precursor computa-
tion and therefore, if the prediction is successful, increases the degree of confidence to
the predictor choice. We selected the size of the spacial grid as a = 400km and the
size of time-step ∆t =half-year. Retrospective analysis was conducted following the
same scheme as in the previous case. In order to reduce the influence of the boundary
conditions the large square covering all the seismic events in the catalog used in the
computations was reduced by two layers of elementary cells from each boundary. As a
result of optimization the forecast information efficiency was 0.549, i.e., the forecasting
algorithm extracted around 55% of all available information about seismic events when
applied to the given catalog.
In the case of binary alarm announcement the space-time share of alarmed cells
was 3.1% and the share of missed targets was 8.3%. This result is comparable with
20 V.Gertsik, M.Kelbert, A.Krichevets
the best forecasts available in the literature and obtained by the other methods (in the
cases when the quantitive parameters of algorithms are presented in the publications).
In the case of forward forecast the two most powerful earthquakes, i.e., SAE and
its major aftershock, happened in the alarm zone of the second level, and two other
events with smaller magnitudes in the fourth alarm zone. Note that in case of binary
alarm announcement one would register a square with 9 elementary cells with only one
alarmed and 8 quiet cells. In this case no reliable forward forecast is possible.
Acknowledgements We would like to thank V.Pisarenko and G.Sobolev for stimulat-
ing discussions that gave us the impulse for writing this paper.
References
Akaike, 1974. Akaike, H., A new look at the statistical model identification,IEEE Trans. Automatic Control, 19, n. 6, 716-723, 1974.
Bolshev, Smirnov, 1965 . Bolshev L. N., Smirnov V. N., Mathematical StatisticalTables (in Russian). V.A. Steklov Matematical Institute,Academy of Sciences, Moscow, USSR, 1965, 464 pp.
Field, 2007. Field, E. H., Overview on the working group for the devel-opment of Regional Earthquake Likelihood Models (RELM),Seismol. Res. Lett. 7â 16, 2007.
Ghertzik, 2008. Gercsik V., Physical concepts of fracture and prediction ofprobabilities of strong earthquakes, Phys. Solid Earth, 44. n.3, 22-39. 2008.
Gercsik, Kelbertl, 2004. Gercsik V., Kelbert M., On comparision of hypothesis tests inBayesian framework without a loss function, Journ. ModernApplied Statistical Methods, 3, n.2, 399-405. 2004.
Harte and Vere-Jones, 2005. Harte D. and Vere-Jones D., The entropy score and its usesin earthquake fore- casting, Pure and Applied Geophysics.162, n. 6-7, 1229-1253, 2005.
Holliday et al., 2007. Holliday J.R., Chien-chih Chen, Tiampo K.F., Rundle J.B.,Turcott D.L. and Donnellan A., A RELM Earthquake fore-cast based on pattern informatics, Seismological ResearchLett., 78, n.1., 97-93, 2007.
Jackson, 1996 . Jackson D. D., Hypothesis testing and earthquake prediction,Proc. Natl. Acad. Sci. USA, 93, 3772-3775, 1996.
Jordan et all., 2011. Jordan T.H., Chen Y.-T., Gasparini P., Madariaga R., MainI., Marzocchi W., Papdopoulos G., Sobolev G., Yamaoka K.and Zschau J., Operational earthquake forecasting: state ofknowledge and guidelines for utilization, Ann. Geophysics,54, n. 4, 2011.
Kagan and Jackson 2000. Kagan Y. Y., and Jackson D.D., Probabilistic forecasting ofearthquakes, Geophysical Journal International, 143, 438â1453, 2000.
Keilis-Borok., 1996. Keilis-Borok V. I., Intermediate-term earthquake prediction,Proceedings of the National Academy of Sciences of theUnited States of America, 93, n. 9, 3748-3755, 1996.
Keilis-Borok, Kossobokov, 1990. Keilis-Borok V.I., Kossobokov V.G. , Premonitory activationof earthquake flow: algorithm M8. Physics of the Earth andPlanetary Interiors, 61. 73-83, 1990.
Kelbert, Suhov, 2013. Kelbert M., Suhov Y., Information Theory and Coding byExample. Cambridge Univ. Press: Cambridge, 2013, 530 pp.
Kolmogorov, 1965 . Kolmogorov A. N., Three approaches to the definition of thenotion of infor- mation amount, Probl. Information Trans-mission, 1, n. 1, 3-11, 1965.
Kolmogorov, 1933. Kolmogorov A. N., On the suitability of statistically obtainedprediction formulas, Zh. Geoz., 3, 7882, 1933.
Earthquake forecasting: Statistics and Information 21
Kolmogorov, 1933a. Kolmogorov A. N., Sulla determinazione empirica di unalegge distribuzione, G. Ist. Ital. Attuari., 4 (1), 83-91, 1933.
Molchan, 1990. Molchan G. M., Strategies in strong earthquake prediction,Phys. Earth Plan. Int., 61, 84-98, 1990.
Molchan and Keilis-Borok, 2008. Molchan, G. M. and Keilis-Borok V. I., Earthquake predic-tion: probabilistic aspect, Geophys. Journ. Int., 173, 1012-1017, 2008.
Pisarenko et al., 2008. Pisarenko V. F., Sornette A., Sornette D., Rodkin M. V.,New approach to the characterization of Mmax and of thetail of the distribution of earthquake magnitudes, Pure andApplied Geophysics, 165, n 5 , 847-888, 2008
Prohorov, Rozanov, 1969. Prohorov Yu.V., Rozanov Yu.A., Probability theory, ba-sic concepts. Limit theorems, random processes, Springer-Verlag, 1969, 401 pp.
Smirnov, 1939. Smirnov N. V., On deviations from the empirical distributioncurve, Mat. Sb., 6(48), 1, 3-24, 1939.
Vere-Jones, 1998. Vere-Jones D., Probability and information gain for earth-quake forecasting, in: Problems of Geodynamics and Seismol-ogy, Computational Seismology, Issue 30, Moscow, 248-263,1998.
WGNCEP, 2011. WGNCEP, 2011 WGNCEP (Working Group on NorthernCalifornia Earthquake Potential), The Uniform CaliforniaEarthquake Rupture Forecast, Version 3 (UCERF3) ProjectPlan : U. S. Geological Survey Open-File Report 1-176, 2011.